`holoviews.element. Curve` does not accept pandas data frame where index is `TimedeltaIndex`

The following code works as expected:

import pandas as pd
import holoviews as hv
hv.extension('bokeh', 'matplotlib')
import numpy as np

dt = pd.date_range('2018-02-23T12:34:45', '2018-02-23T12:47:00', freq='1s', name='t')
data = pd.DataFrame(np.sin((dt-dt[0]) / np.timedelta64(1,'m')).values, columns=('A', ), index=dt)

hv.Curve(data)

but

data.index = dt - dt[0]
hv.Curve(data)

raises an error:

TypeError: ufunc 'isfinite' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''

I would like to plot pandas data frames indexed both by DatetimeIndex (works) or TimedeltaIndex (does not work)

[philippjfr]

The formatting issues are subsumed by https://github.com/holoviz/holoviews/issues/3143 so I’ll close this now.

[miccoli]

I’ve done some testing using my patched bokeh miccoli/bokeh@0545753. Starting from numpy datetime64 and timedelta64 types, I’ve created two different pandas data frames.

dt = np.arange('2018-03-14T15:30:21', '2020-06-28T13:30:23', dtype='datetime64[h]')
td = dt - dt[len(dt)//2]
y = np.sin(4*np.pi*td / td.ptp())

pddt = pd.DataFrame(y, index=dt, columns=['sin'])
pddt.index.name = 'datetime'

pdtd = pd.DataFrame(y, index=td, columns=['sin'])
pdtd.index.name = 'timedelta'

Here the results of hv.Curve(pddt) and hv.Curve(pdtd)

The problem here is that in bokeh timedelta are scaled as ms (millisecond) and represented with a plain linear axis. This is could be OK, but the choice of ms is not very sensible in my opinion… Using the matplotlib backend one gets the same linear axis but with timedelta scaled as ns (nanosecond).

Things are not that better in pandas: pdtd.plot(figsize=(14,3), grid=True)

Things to fix, in my opinion:

1. consistent behaviour for bokeh and matplotlib
2. if a linear axis is to be used (without dynamic scaling like the bokeh datetime axis), times should be given in seconds and properly labelled as so.

Of course the ideal solution would be to implement native timedelta support in bokeh (extending the datetime axis or maybe creating a new timedelta axis type.)